Cold User Similarity Algorithms Based On Collaborative Filtering

With the rapid development of Internet technology and e-commerce,the type and quantity of Internet service information is increasing at a frightening pace which has resulted in the inevitable phe nomenon of information overload.Recommender systems(RSs)are emerging in response to this challenges.Collaborative filtering(CF)algorithm is the core of the RS which use the historical track record of users' activities and possibly personal profiles to uncover their preferences and tastes,to help them find products or services of interest from a large volume of available choices,and to provide users with high quality personalized recommendations.Nevertheless,CF is still confronted with a lot of issues in practical applications.Cold start and data sparsity are the key problems in CF recommendation technology that have not been solved effectively.The existing user similarity algorithms based on CF only use rating values in the rating matrix and ignore the influence of preference difference of co-ratings,users' personal rating preferences and popular items on the similarity between users when calculating the similarities between cold start users and other users.In this case,the accuracy of user similarity obtained will be greatly reduced,so it is difficult to accurately and efficiently predict the interest of the target user,resulting in low accuracy of the recommendation results based on the CF recommendation algorithm.In this paper,we detailed ly analyze some problems of user similarity algorithm based on CF in dealing with the problem of new user cold start and data sparsity,put forward some improvement ideas and have achieved some research results.The main contents and innovation points can be summarized as follows:1)We propose a heuristic similarity algorithm considering the preference difference of co-ratings between users.The new algorithm is based on idea of the PIP and MJD similarity measures and calculates similarity based on the difference of co-ratings for each pair users.First,it calculates the preference weights of co-ratings using the proportion the difference of co-ratings.On the basis of this,three kinds of impact factors: Proximity,Impact and Popularity are calculated.Finally,we will be able to obtain a global improved similarity measure between users.The new algorithm takes into account both domain specific meanings of data and the preference difference of co-ratings for each pair users simultaneously,which effectively avoids the unreasonable increase of the similarity and improves the discrimination degree of similar users.Furthermore,the weight calculation of different preference is simple and not time consuming.2)We propose a user similarity algorithm considering user s' personal rating difference and popular items.The new similarity algorithm is composed of three similarity factors(PMSD,SD and Preference),which takes into account the influence of popular items on the similarity between users in a particular dataset and combining it with the Mean Squared Difference(MSD),utilizes all available information of users' ratings including numerical and non-numerical information of users' ratings and expresses different characteristics.What's more,the new similarity algorithm also considers the preference difference of co-ratings for each pair users by giving different penalty values based on the different preferences of co-ratings.Furthermore,the mean or/and variance of user rating also be adopted to reflect the rating preference of any user.We carry out experiments to test and verify the performance of the two new similarity algorithms.More comparisons with other traditional and improved user similarity algorithms are conducted.New similarity measures can indicate better recommendation performance and prediction accuracy in cold start conditions.Both the experiment results and theoretical analysis show that the proposed algorithms can obtain better recommendation performance on MAE,coverage,precision and recall,and greatly improves the prediction accuracy of CF algorithm and the recommendation quality of RS under both new user cold start and data sparsity conditions.